He says impractical mixture-of-deltas solutions do not invalidate other derivations.
@torchcompiled It isn't useless and it isn't crazy
Any maximum likelihood method has a mixture of deltas solution, that one is actually useless
It’s crazy that the diffusion model has a closed-form solution that’s effectively useless and only capable of reproducing exact memorizations.
It’s the imperfection that lets them work, “failing gracefully” if you will
Users affirm the value of closed-form diffusion model solutions because empirical scores generalized effectively over a large part of the time range.
@LucaAmb It is crazy to me even if it has a logical reasoning :3
@torchcompiled It isn't useless and it isn't crazy
Any maximum likelihood method has a mixture of deltas solution, that one is actually useless
@torchcompiled Alright, it's not useless though! The empirical score generalized for a large part of the time range!
He says impractical mixture-of-deltas solutions do not invalidate other derivations.
@torchcompiled It isn't useless and it isn't crazy
Any maximum likelihood method has a mixture of deltas solution, that one is actually useless
It’s crazy that the diffusion model has a closed-form solution that’s effectively useless and only capable of reproducing exact memorizations.
It’s the imperfection that lets them work, “failing gracefully” if you will